The Pentagon’s top research agency is taking a no-nonsense approach to advancing development of a fundamental component of AI technology known as “machine common sense.”
The Defense Advanced Research Projects Agency (DARPA), which announced a multi-year $2 billion “AI Next” campaign in September, is tightening its focus on developing machine “common sense” reasoning. The basic research effort includes two technology tracks, the first focusing on fundamentals, or “learning like a child.”
The second includes developing models based on deep learning and other approaches to develop frameworks that can solve problems based on current AI industry benchmarks. The effort also provides momentum for academic research in this esoteric but vital area of cognitive studies that dates to the 1960s.
The DARPA program starts with the premise the machines lack basic knowledge about the physical world that all humans possess. Expressing and encoding that capability has so far confounded AI researchers who tend to focus on either narrow or general AI problems. Those models are considered “brittle,” falling well short of human cognition.
“It’s new territory [and] it turns out to be incredibly difficult,” said David Gunning, a program manager in DARPA’s Information Innovation Office. “Every [human] action is on top of this iceberg of common-sense knowledge.” Even a one-year-old child “has an incredible understanding of the physical world” that machines can’t come close to matching, Gunning said.
If the basic research program pans out, DARPA hopes to use the resulting models to develop new machine learning applications. “We’ll get something started that will synergize” further industry development, Gunning added in an interview.
Potential common sense “services” include models that learn from experience as might a child learning spatial navigation. Another would mimic a research librarian browsing the web to gather facts that could be used to provide common-sense answers to natural language or image-based questions about the physical world.
One potential application dubbed “sensemaking” would augment AI systems that analyze sensor data. A machine common sense model could then help interpret and understand real-world scenarios. Another described by Gunning is a “simulated common sense agent [operating] in a 3-D simulated environment.”
Gunning’s office kicked the effort with a recent “ Proposers’ Day ” designed to brief potential bidders on program requirements. The agency said it would fund research exploring developmental psychology, then establish a set of cognitive development milestones for determining how the resulting computational models learn in three areas: experience learning, prediction and “expectation,” as well as problem solving.